TOPOLOGY OPTIMIZATION BY THE HOMOGENIZATION - CMAP

1

OPTIMAL DESIGN OF STRUCTURES (MAP 562)

G. ALLAIRE

February 11th, 2015

Department of Applied Mathematics, Ecole Polytechnique

CHAPTER VII (continued)

TOPOLOGY OPTIMIZATION

BY THE HOMOGENIZATION METHOD

G. Allaire, Ecole Polytechnique Optimal design of structures

2

7.5 Shape optimization in the elasticity setting

Very similar to the conductivity setting but there are some additional hurdles.

We shall review the results without proofs.

The basic ingredients of the homogenization method are the sames:

introduction of composite designs characterized by (θ, A∗),

Hashin-Shtrikman bounds for composites,

sequential laminates are optimal microstructures.


3

Model problem: compliance minimization

ΓD

N

Ω

Γ

D

Γ

Bounded working domain D ∈ IRN (N = 2, 3).

Linear isotropic elastic material, with Hooke’s law A

A = (κ− 2µ

N)I2 ⊗ I2 + 2µI4, 0 < κ, µ < +∞

Admissible shape = subset Ω ⊂ D.


4

Boundary ∂Ω = Γ ∪ ΓN ∪ ΓD with ΓN and ΓD fixed.

− divσ = 0 in Ω

σ = 2µe(u) + λ tr(e(u)) Id in Ω

u = 0 on ΓD

σn = g on ΓN

σn = 0 on Γ,

Weight is minimized and rigidity is maximized. Let ℓ > 0 be a Lagrange

multiplier, the objective function is

infΩ⊂D

J(Ω) =

∫

ΓN

g · u ds+ ℓ

∫

Ω

dx

.


5

This shape optimization problem can be approximated by a two-phase

optimization problem: the original material A and the holes of rigidity B ≈ 0.

The Hooke’s law of the mixture in D is

χΩ(x)A+(

1− χΩ(x))

B ≈ χΩ(x)A

The admissible set is

Uad =

χ ∈ L∞ (D; 0, 1)

.

As in conductivity/membrane case we can apply the relaxation approach

based on homogenization theory.

The homogenization method can be generalized to the elasticity setting.


6

Homogenized formulation of shape optimization

We introduce composite structures characterized by a local volume fraction

θ(x) of the phase A (taking any values in the range [0, 1]) and an homogenized

tensor A∗(x), corresponding to its microstructure.

The set of admissible homogenized designs is

U∗ad =

(θ, A∗) ∈ L∞(

D; [0, 1]× IRN4)

, A∗(x) ∈ Gθ(x) in D

.

The homogenized state equation is

σ = A∗e(u) with e(u) = 12 (∇u+ (∇u)t) ,

divσ = 0 in D,

u = 0 on ΓD

σn = g on ΓN

σn = 0 on ∂D \ (ΓD ∪ ΓN ).


7

The homogenized compliance is defined by

c(θ, A∗) =

∫

ΓN

g · u ds.

The relaxed or homogenized optimization problem is

min(θ,A∗)∈U∗

ad

J(θ, A∗) = c(θ, A∗) + ℓ

∫

D

θ(x) dx

.

Major inconvenient: in the elasticity setting an explicit characterization of Gθ

is still lacking !

Fortunately, for compliance one can replace Gθ by its explicit subset of

laminated composites.


8

The key argument to avoid the knowledge of Gθ is that, thanks to the

complementary energy minimization, compliance can be rewritten as

c(θ, A∗) =

∫

ΓN

g · u ds = mindivσ=0 in Dσn=g on ΓN

σn=0 on ∂D\ΓN∪ΓD

∫

D

A∗−1σ · σ dx.

The shape optimization problem thus becomes a double minimization (we

already used this argument in chapter 5).


9

Exchanging the order of minimizations

The shape optimization problem is

min(θ,A⋆)∈U⋆

ad

mindivσ=0 in Dσn=g on ΓN


∫

D

A∗−1σ · σ dx+ ℓ

∫

D

θ(x) dx

.

Since the order of minimization is irrelevent, it can be rewritten



min(θ,A⋆)∈U⋆

ad

∫

D

A∗−1σ · σ dx+ ℓ

∫

D

θ(x) dx

.

The minimization with respect to the design parameters (θ, A∗) is local, thus



∫

D

min0≤θ≤1A∗∈Gθ

(

A∗−1σ · σ + ℓθ)

(x) dx.


10

For a given stress tensor σ, the minimization of complementary energy

minA∗∈Gθ

A∗−1σ · σ

is a classical problem in homogenization, of finding optimal bounds on the

effective properties of composite materials.

It turns out that we can restrict ourselves to sequential laminates which form

an explicit subset Lθ of Gθ.

Such a simplification is made possible because compliance is the objective

function.


11

7.5.2 Sequential laminates in elasticity

Aξ = 2µAξ + λA(trξ)I, Bξ = 2µBξ + λB(trξ)I,

with the identity matrix I2, and κA,B = λA,B + 2µA,B/N . We assume B to be

weaker than A

0 ≤ µB < µA, 0 ≤ κB < κA.

We work with stresses rather than strains, thus we use inverse elasticity

tensors.

Lemma 7.24. The Hooke’s law of a simple laminate of A and B in

proportions θ and (1− θ), respectively, in the direction e, is

(1− θ)(

A∗−1 −A−1)−1

=(

B−1 −A−1)−1

+ θf cA(e)

with f cA(e) the tensor defined, for any symmetric matrix ξ, by

f cA(ei)ξ · ξ = Aξ · ξ − 1

µA

|Aξei|2 +µA + λA

µA(2µA + λA)((Aξ)ei · ei)2.


12

Reiterated lamination formula

Proposition 7.25. A rank-p sequential laminate with matrix A and

inclusions B, in proportions θ and (1− θ), respectively, in the directions

(ei)1≤i≤p with parameters (mi)1≤i≤p such that 0 ≤ mi ≤ 1 and∑p

i=1 mi = 1,

is given by

(1− θ)(

A∗−1 −A−1)−1

=(

B−1 −A−1)−1

+ θ

p∑

i=1

mifcA(ei)

1ε

1εε2 >>

e2

Α =

Β =e1


13

7.5.3 Hashin-Shtrikman bounds in elasticity

Theorem 7.26. Let A∗ be a homogenized elasticity tensor in Gθ which is

assumed isotropic

A∗ = 2µ∗I4 +

(

κ∗ −2µ∗

N

)

I2 ⊗ I2.

Its bulk κ∗ and shear µ∗ moduli satisfy

1− θ

κA − κ∗≤ 1

κA − κB

+θ

2µA + λA

andθ

κ∗ − κB

≤ 1

κA − κB

+1− θ

2µB + λB

1− θ

2(µA − µ∗)≤ 1

2(µA − µB)+

θ(N − 1)(κA + 2µA)

(N2 +N − 2)µA(2µA + λA)

θ

2(µ∗ − µB)≤ 1

2(µA − µB)− (1− θ)(N − 1)(κB + 2µB)

(N2 +N − 2)µB(2µB + λB).

Furthermore, the two lower bounds, as well as the two upper bounds are

simultaneously attained by a rank-p sequential laminate with p = 3 if N = 2,

and p = 6 if N = 3.


14

Hashin-Shtrikman bounds in elasticity

κ

µ

θ+µ

µθ−

κ θ− κ θ

+


15

Proposition 7.27. Let Gθ be the set of all homogenized elasticity tensors

obtained by mixing the two phases A and B in proportions θ and (1− θ). Let

Lθ be the subset of Gθ made of sequential laminated composites. For any

stress σ,

HS(σ) = minA∗∈Gθ

A∗−1σ · σ = minA∗∈Lθ

A∗−1σ · σ.

Furthermore, the minimum is attained by a rank-N sequential laminate with

lamination directions given by the eigendirections of σ.

Remark.

An optimal tensor A∗ can be interpreted as the most rigid composite

material in Gθ able to sustain the stress σ.

HS(σ) is called Hashin-Shtrikman optimal energy bound.

In the conductivity setting, a rank-1 laminate was enough...

Practical conclusion: Gθ can be replaced by Lθ.


16

Explicit computation of the optimal bound

When B = 0 one can obtain an explicit formula for the bound:

minA∗∈Gθ

A∗−1σ · σ = HS(σ) = A−1σ · σ +1− θ

θg∗(σ)

2-D case.

g∗(σ) =κ+ µ

4µκ(|σ1|+ |σ2|)2

with σ1, σ2 the eigenvalues of σ. Furthermore, an optimal rank-2 sequential

laminate is given by the parameters

m1 =|σ2|

|σ1|+ |σ2|, m2 =

|σ1||σ1|+ |σ2|

.


17

3-D simplified case with λA = 0. We label the eigenvalues of σ as

|σ1| ≤ |σ2| ≤ |σ3|.

g∗(σ) =1

4µ

(|σ1|+ |σ2|+ |σ3|)2 if |σ3| ≤ |σ1|+ |σ2|

2(

(|σ1|+ |σ2|)2 + |σ3|2)

if |σ3| ≥ |σ1|+ |σ2|

In the first regime, an optimal rank-3 sequential laminate is given by

m1 =|σ3|+ |σ2| − |σ1||σ1|+ |σ2|+ |σ3|

, m2 =|σ1| − |σ2|+ |σ3||σ1|+ |σ2|+ |σ3|

, m3 =|σ1|+ |σ2| − |σ3||σ1|+ |σ2|+ |σ3|

,

and in the second regime, an optimal rank-2 sequential laminate is

m1 =|σ2|

|σ1|+ |σ2|, m2 =

|σ1||σ1|+ |σ2|

, m3 = 0.

(General 3-D case known but more complicated.)


18

7.5.4 Homogenized formulation of shape optimization



∫

D

min0≤θ≤1A∗∈Gθ

(

A∗−1σ · σ + ℓθ)

dx.

Optimality condition. If (θ, A∗, σ) is a minimizer, then A∗ is a rank-N

sequential laminate aligned with σ and with explicit proportions

A∗−1 = A−1 +1− θ

θ

(

N∑

i=1

mifcA(ei)

)−1

,

and θ is given in 2-D (similar formula in 3-D)

θopt = min

(

1,

√

κ+ µ

4µκℓ(|σ1|+ |σ2|)

)

,

where σ is the solution of the homogenized equation.


19

Existence theory

Original shape optimization problem

infΩ⊂D

J(Ω) =

∫

ΓN

g · u ds+ ℓ

∫

Ω

dx. (1)

Homogenized (or relaxed) formulation of the problem

minA∗∈Gθ

0≤θ≤1

J(θ, A∗) =

∫

ΓN

g · u ds+ ℓ

∫

D

θ dx. (2)

Theorem 7.30. The homogenized formulation (2) is the relaxation of the

original problem (1) in the sense where

1. there exists, at least, one optimal composite shape (θ, A∗) minimizing (2),

2. any minimizing sequence of classical shapes Ω for (1) converges, in the

sense of homogenization, to a minimizer (θ, A∗) of (2),

3. the minimal values of the original and homogenized objective functions

coincide.


20

7.5.5 Numerical algorithm

Double “alternating” minimization in σ and in (θ, A∗).

• intialization of the shape (θ0, A∗0)

• iterations n ≥ 1 until convergence

– given a shape (θn−1, A∗n−1), we compute the stress σn by solving a

linear elasticity problem (by a finite element method)

– given a stress field σn, we update the new design parameters (θn, A∗n)

with the explicit optimality formula in terms of σn.

Remarks.

For compliance, the problem is self-adjoint.

Micro-macro method (local microstructure / global density).


21

Remarks

The objective function always decreases.

Algorithm of the type “optimality criteria”.

Algorithme of “shape capturing” on a fixed mesh of Ω.

We replace void by a weak “ersatz” material, or we impose θ ≥ 10−3 to

get an invertible rigidity matrix.

A few tens of iterations are sufficient to converge.


22

Example: optimal cantilever


23

Penalization

The previous algorithm compute composite shapes instead of classical

shapes.

We thus use a penalization technique to force the density in taking values

close to 0 or 1.

Algorithm: after convergence to a composite shape, we perform a few more

iterations with a penalized density

θpen =1− cos(πθopt)

2.

If 0 < θopt < 1/2, then θpen < θopt, while, if 1/2 < θopt < 1, then θpen > θopt.


24


25


26

Convergence history:

objective function (left), and residual (right),

in terms of the iteration number.

iteration number

obje

ctiv

e fu

ncti

on

0 10050

0.4

0.5

0.6

0.7

0.8

0.9

1

1.1

iteration number

conv

erge

nce

crit

erio

n

0 10050-510

-410

-310

-210

-110

010


27

Example: optimal bridge


28

7.5.6. Convexification and “fictitious materials”

Idea. In the homogenization method composite materials are introduced but

discarded at the end by penalization. Can we simplify the approach by

introducing merely a density θ ?

A classical shape is parametrized by χ(x) ∈ 0, 1.If we convexify this admissible set, we obtain θ(x) ∈ [0, 1].

The Hooke’s law, which was χ(x)A, becomes θ(x)A. We also call this

fictitious materials because one can not realize them by a true

homogenization process (in general). Combined with a penalization scheme,

this methode is called SIMP (Solid Isotropic Material with Penalization).


29

Convexified formulation with 0 ≤ θ(x) ≤ 1

σ = θ(x)Ae(u) with e(u) = 12 (∇u+ (∇u)t) ,

divσ = 0 in D,

u = 0 on ΓD

σn = g on ΓN

σn = 0 on ∂D \ (ΓD ∪ ΓN ).

Compliance minimization

min0≤θ(x)≤1

(

c(θ) + ℓ

∫

D

θ(x)

)

.

with

c(θ) =

∫

ΓN

g · u =

∫

D

(θ(x)A)−1σ · σ = mindivτ=0 in Dτn=g on ΓN

τn=0 on ∂D\ΓN∪ΓD

∫

D

(θ(x)A)−1τ · τ dx.

Now, there is only one single design parameter: the material density θ (the

microstructure A∗ has disappeared).


30

Existence of solutions

Theorem 7.33. The convexified formulation

min0≤θ(x)≤1

mindivτ=0 in Dτn=g on ΓN

τn=0 on ∂D\ΓN∪ΓD

∫

D

(θ(x)A)−1τ · τ dx+ ℓ

∫

D

θ dx

admits at least one solution.

Proof. The function, defined on IR+ ×Msn,

φ(a, σ) = a−1A−1σ · σ,

is convex because

φ(a, σ) = φ(a0, σ0) +Dφ(a0, σ0) · (a− a0, σ − σ0) + φ(a, σ − aa−10 σ0),

where the derivative Dφ is given by

Dφ(a0, σ0) · (b, τ) = − b

a20A−1σ0 · σ0 + 2a−1

0 A−1σ0 · τ.


31

Optimality condition

If we exchange the minimizations in τ and in θ, we can compute the optimal θ

which is

θ(x) =

1 if A−1τ · τ ≥ ℓ√ℓ−1A−1τ · τ if A−1τ · τ ≤ ℓ

Again we can use an “alternating” double minimization algorithm.


32

Numerical algorithm

• intialization of the shape θ0

• iterations k ≥ 1 until convergence

– given a shape θk−1, we compute the stress σk by solving an elasticity

problem (by a finite element method)

– given a stress field σk, we update the new material density θk with the

explicit optimality formula in terms of σk.

Penalization: we use a penalized density

θpen =1− cos(πθopt)

2or (SIMP) θpen = θp p > 1.

In practice: it is extremely simple ! But the numerical results are not as

good ! An explanation is the lack of a relaxation theorem.

Be careful: very delicate monitoring of the penalization...


33

Optimal bridge by the convexification method

iteration number

com

plia

nce

0 10050

0.2

0.3

0.4

0.15

0.25

0.35

0.45

convexificationhomogenization


34

Conclusion

SIMP (or convexification, or “fictitious materials”) is very simple and

very popular (many commercial codes are using it).

SIMP uses very few informations on composites !

On the contrary to the homogenization method, SIMP is not a

relaxation method: it changes the problem !

There is a gap between the true minimal value of the objective function

and that of SIMP.

SIMP can be delicate to monitor: how to increase the penalization

parameter ?


35

Generalizations of the homogenization method

multiple loads

vibration eigenfrequency

general criterion of the least square type

The two first cases are self-adjoint and we have a complete understanding and

justification of the relaxation process. However, the third case is not

self-adjoint and only a partial relaxation is known.


36

Multiple loads

For n loads (fi)1≤i≤n, the homogenized formulation is

mindivσi=0 in Dσin=gi on ΓN

∫

D

min0≤θ≤1

minA∗∈Lθ

(

n∑

i=1

A∗−1σi · σi + ℓθ

)

dx

with A∗ ∈ Lθ and

(1− θ)(

A∗−1 −A−1)−1

=(

B−1 −A−1)−1

+ θ

p∑

i=1

mifcA(ei)

The optimal laminate is no more of rank N . The mi’s optimization is now

done numerically (with numerous enough lamination directions).


37

Optimal bridge for 3 simultaneously applied loads


38

Optimal bridge for 3 independently applied loads


39

Vibration eigenfrequencies

We maximize the first vibration eigenfrequency

ω21(θ, A

∗) = minu∈H

∫

D

A∗e(u) · e(u)dx∫

D

ρ|u|2dx.

with the density ρ = θρA + (1− θ)ρB , and the space of admissible

displacements H =

u ∈ H1(D)N such that u = 0 on ΓD

.

The homogenized formulation is

max0≤θ≤1

minu∈H

∫

D

(

maxA∗∈Lθ

A∗e(u) · e(u))

dx∫

D

ρ|u|2dx+ ℓ

∫

D

θ(x)dx

,

with Lθ the set of sequential laminates.

Be careful: there is a max-min which can not be exchanged...


40


41

Least square objective functions

Classical two-phase formulation:

infχ∈L∞(Ω;0,1)

J(χ) =

∫

Ω

k(x)|uχ(x)− u0(x)|2 dx+ ℓ

∫

Ω

χ(x)dx

where uχ is solution of

− div (Aχe(uχ)) = f in Ω

uχ = 0 on ∂Ω,

with a Hooke’s law Aχ = χA+ (1− χ)B.


42

Homogenized formulation:

min(θ,A∗)

J∗(θ, A∗) =

∫

Ω

(

k|u− u0|2 + ℓθ)

dx

with u solution of

− div (A∗e(u)) = f in Ω

u = 0 on ∂Ω,

Difficulty: we don’t know Gθ and we cannot replace it by Lθ. In other

words, we don’t know which microstructures are optimal...

Partial relaxation: we nevertheless replace Gθ by Lθ. We thus loose the

existence of an optimal solution but we keep the link with the original

problem.


43

Partial relaxation

We restrict ourselves to sequential laminates A∗ with matrix A and inclusions

B. The number of laminations and their directions are fixed. We merely

optimize with respect to θ and the proportions (mi)1≤i≤p

(1− θ) (A−A∗)−1

= (A−B)−1 − θ

q∑

i=1

mifA(ei),

with ∀e ∈ IRN , |e| = 1, ∀ξ symmetric matrix

fA(e)ξ · ξ =1

µA

(

|ξe|2 − (ξe · e)2)

+1

λA + 2µA

(ξe · e)2.

Thus, the objective function is

J∗(θ, A∗) ≡ J∗(θ,mi)

with the constraints 0 ≤ θ ≤ 1, mi ≥ 0,∑p

i=1mi = 1.

We compute its gradient with the help of an adjoint state.


44

Adjoint state

Typical example of an objective function

J∗(θ, A∗) =

∫

Ω

k(x)|u(x)− u0(x)|2dx+ ℓ

∫

Ω

θ dx

Adjoint state

− div (A∗e(p)) = 2k(x)(u(x)− u0(x)) in Ω

p = 0 on ∂Ω


45

Gradient

∇θJ∗(x) = ℓ+

∂A∗

∂θe(u) · e(p),

∇miJ∗(x) =

∂A∗

∂mi

(x)e(u) · e(p),

and

∂A∗

∂θ(x) = T−1

(

(A−B)−1 −q∑

i=1

mifA(ei)

)

T−1,

∂A∗

∂mi

(x) = −θ(1− θ)T−1fA(ei)T−1,

T = (A−B)−1 − θ

q∑

i=1

mifA(ei) .


46

Numerical algorithm of gradient type

Projected gradient with a variable step:

1. Initialization of the design parameters θ0,mi,0 (for example, constants

satisfying the constraints).

2. Iterations until convergence, for k ≥ 0:

(a) Computation of the state uk and the adjoint pk, with the previous

design parameters θk,mi,k.

(b) Update of the design parameters :

θk+1 = max (0,min (1, θk − tk∇θJ∗k )) ,

mi,k+1 = max (0,mi,k − tk∇miJ∗k + ℓk) ,

where ℓk is a Lagrange multiplier for the constraint∑q

i=1mi,k = 1, iteratively

updated, and tk > 0 is a descent step such that J∗(θk+1,mk+1) < J∗(θk,mk).


47

Example: force inverter

C(x) = 0

F

C(x) = 1


48

Other methods of topology optimization

Discrete 0/1 optimization: genetic algorithms.

Level set methods based on geometric optimization.

Topological derivative: sensitivity to the nucleation of a small hole.

Phase-field methods.


49

Commercial softwares and industrial applications

See the web page:

http://www.cmap.polytechnique.fr/~optopo/links.html


50

Industrial applications

Automotive industry.

Aerospace industry.

Civil engineering, architecture.

Nano-technologies, MEMS.

Optics, wave guides.


TOPOLOGY OPTIMIZATION BY THE HOMOGENIZATION - CMAP

Documents

Transcript of TOPOLOGY OPTIMIZATION BY THE HOMOGENIZATION - CMAP